The model, already very potent in standard Croatian, was finetuned for 80 epochs with an effective batch size of 16. Performance was inspected every 4 epochs, and the latest checkpoint is uploaded here. Character error rate has been brought down from 11.54% to 3.95%, while word error rate has been lowered from 35.43% to 16.83%.

Developed by: Nikola Ljubešić, Peter Rupnik, Tea Perinčić
Language(s) (NLP): Croatian (hrv) - Chakavian dialect (ckm)
License: Creative Commons - Share Alike 4.0
Finetuned from model: openai/whisper-large-v3

Model Sources

Repository: GitHub
Paper: Coming soon
Dataset: Mići Princ

Example use:

python

import torch
from datasets import load_dataset
from transformers import AutoModelForSpeechSeq2Seq, AutoProcessor, pipeline
from transformers.pipelines.pt_utils import KeyDataset

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model_id = "classla/whisper-large-v3-mici-princ"
model = AutoModelForSpeechSeq2Seq.from_pretrained(
    model_id,
)

model.to(device)
processor = AutoProcessor.from_pretrained(model_id)

ds = load_dataset("classla/Mici_Princ", split="test")
pipe = pipeline(
    "automatic-speech-recognition",
    model=model,
    tokenizer=processor.tokenizer,
    feature_extractor=processor.feature_extractor,
    max_new_tokens=128,
    chunk_length_s=30,
    batch_size=16,
    return_timestamps=True,
    device=device,
)

result = pipe(
    KeyDataset(ds, "audio"),
    generate_kwargs={"language": "croatian"},
)

for i in result:
    print(i)

# Output:
# {'text': ' Šesti planet je biv deset put veći. Na njin je bivav niki stari čovik ki je pisav vele knjige.', 'chunks': [{'timestamp': (0.0, 7.18), 'text': ' Šesti planet je biv deset put veći. Na njin je bivav niki stari čovik ki je pisav vele knjige.'}]}
# ...

Training Details

Preprocessing

Model was trained on the normalized_text attribute of the Mići Princ dataset. This means that the data included capital letters and punctuation, except bullet points, newlines, and quotation marks. Special characters, present in the dialect, but not in standard Croatian, were substituted.

Only the train split was used in training.

Training Hyperparameters

    per_device_train_batch_size=4,
    gradient_accumulation_steps=4,
    learning_rate=1e-5,
    warmup_steps=100,
    max_steps=277 * 80,
    gradient_checkpointing=True,
    predict_with_generate=True,
    generation_max_length=225,
    save_steps=277,

Evaluation

For evaluation, the test split of the Mići Princ dataset was used. The test split consists of two known speakers, Autor and Mići Princ, and two unknown speakers, Geograf and Dilavac. Important to note is that each speaker uses a different micro-dialect, so the test data is challenging on including two new micro-dialects.

Metrics

speaker	WER vanilla	WER fine-tuned	WER reduction	CER vanilla	CER fine-tuned	CER reduction
all	35.43%	16.83%	52.50%	11.54%	3.95%	65.77%
Autor	38.96%	14.29%	63.32%	10.24%	2.93%	71.39%
Geograf	20.94%	11.57%	44.75%	4.99%	2.19%	56.11%
Mići Princ	45.32%	16.62%	63.33%	12.21%	5.09%	58.31%
Dilavac	39.60%	23.70%	40.15%	18.55%	5.27%	71.59%

Citation

Coming soon.

Model Card Authors

Peter Rupnik
Nikola Ljubešić

Model Card Contact

https://huggingface.co/5roop